Learning robot manipulation from human videos alone, no teleoperation demos needed
burkov · x · 2026-10-13
Robotics faces a scaling bottleneck: teleoperated demos are expensive and hardware-specific, while ordinary human videos suffer from embodiment mismatch and joint-limit/collision failures. This piece evaluates a system that learns reusable object manipulation skills entirely from human video demonstrations and adapts them to varied robots and environments — with zero real-world robot demos in target tasks.
More from Embodied
- Ex-Apple engineer eyeing Surface with NVIDIA RTX Spark: 128GB + CUDA in a backpack — LinusEkenstam · 2026-10-13
- Baidu's Apollo Go opens public robotaxi rides in Hong Kong's Tung Chung from Oct 21 — Baidu_Inc · 2026-10-13
- Google's AgentHands adds spatially grounded, synchronized hand gestures to XR AI agents — CurieuxExplorer · 2026-10-13
- Nature Will launches Natura-1, a foundation model for the physical world trained on 1M hours of egocentric video — testingcatalog · 2026-10-13
- Boston Dynamics Cuts the Pinky: New Atlas Hand Has 13 DoF, Targets 100K Units a Year — lukas_m_ziegler · 2026-10-13
- Meta AI glasses proactively flagged two gate changes during travel, outpacing airline alerts — NinaDSchick · 2026-10-13